In the field of computational chemistry, the construction of databases of transition metal complexes (TMCs) from experimental crystal structures is often hindered by the intrinsic complexity of their electronic structure. Although crystallographic databases provide a vast collection of structures exhibiting broad chemical diversity, they generally lack certain key pieces of information required for quantum chemical computations, such as the total molecular charge or the ground-state spin. Beyond these challenges related to the curation and interpretation of experimental data, the statistical prediction of TMC properties remains limited, both by the range of machine learning (ML) models explored and by the diversity of coordination environments and electronic characteristics considered.
This thesis addresses these challenges by developing automated workflows that leverage crystallographic data for computational chemistry purposes. We first present cell2mol, a Python software package designed for the interpretation of molecular crystal structures, with particular emphasis on systems containing TMCs. This tool enables the extraction of molecular connectivity, the assignment of metal oxidation states, and the determination of total molecular charges and bond orders. It has thus enabled the curation of a chemically diverse database comprising more than 31,000 mononuclear TMCs. Building upon this dataset, we develop a statistical model capable of predicting the ground-state spin of 3d TMCs directly from crystal structures with high accuracy. An enhanced version, cell2mol 2.0, further improves performance, extends functionality, and introduces a web interface to facilitate accessibility.
Leveraging an expanded and carefully curated dataset generated with cell2mol, this thesis subsequently trains physics-inspired ML models to predict a variety of TMC properties across a broad range of coordination environments, charge states, and spin states. Finally, by constructing a dataset of problematic multiconfigurational 3d TMCs inspired by previous studies, we analyze multireference diagnostics by reassessing commonly used thresholds, with the goal of training geometric deep learning models capable of predicting these diagnostics.
Overall, this thesis establishes an automated and integrated framework that strengthens electronic structure workflows for TMCs derived from crystal structures while accelerating their large-scale exploration and design. More broadly, it addresses a practical limitation of previous data-driven approaches and significantly expands the chemical space of TMCs accessible to computational investigation.
